[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: IDL 5.3 Performance ?



Richard Tyc wrote:

> Q. for JD Smith: is the is2_53.sav test program alot different than the
> version for IDL 5.2 ?
> 

This is exactly the same program, just "recompiled" for 5.3.  It is very basic,
and does little more than a bunch of housekeeping and stuff to auto-email the
results.  The actual testing is being done with:

time_test3,/NOFILEIO

graphics_times3

and the extracted IO test from time_test3.  When I compare time_test.pro from
IDL5.3 to IDL5.2.1, the *only* differences are the addition of the new compiler
directives:

COMPILE_OPT hidden

to hide the compilation.  The file time_test3.pro differs only in the version
number listed at the top (1.1 vs. 1.1.2.1), but otherwise has *no* differences. 
The story is the same with graphics_times*.pro.

I rechecked the excerpted IO test, and it still checks out.  


> Any comments/ideas ?

The limited 5.3 testing I've done does indicate a slowdown, although these tests
were confined to a few machines only.  For instance, on the chart on the
IDLSpecII page, you can see I've entered my machine (Dell Dimension 400Mhz PII)
twice, and the IO was hardest hit (by far).  To really see what the slow down
is, or if it is general, we need people to recontribute 5.3 idlspec2 entries
from machines they had sent in before, making sure to use exactly the same
descriptions and text so that i can sort those out together.  Just take a look
at http://www.astro.cornell.edu/idlspec/is2_sorted.html to find your old entry.  


In the interest of a little more certainty, I did a side by side time_test3 with
v5.2.1 and v5.3.  I got:

   5.2.1            5.3
 1     0.21400   1     0.20214
 2     0.23642   2     0.23430
 3     0.25929   3     0.26889
 4     0.24083   4     0.25496
 5     0.14772   5     0.19550
 6      1.0328   6      1.0214
 7     0.50235   7     0.50162
 8     0.46114   8     0.46131
 9     0.44494   9     0.44587
10     0.87304  10     0.87744
11     0.67848  11     0.67851
12     0.33605  12     0.24401
13     0.40520  13     0.21952
14     0.27451  14    0.085009
15     0.26790  15     0.25382
16     0.36092  16     0.41647
17      1.4182  17      1.4306
18     0.27534  18     0.26124
19     0.45899  19     0.46114
20     0.46100  20     0.45048
21     0.32254  21     0.32184
22     0.13546  22     0.13159
23     0.90981  23     0.91101
10.7171=Total   10.3288=Total 


So this doesn't really support a slowdown. Test 5 "Mult 512 by 512 byte by
constant and store, 30 times", is marginally slower.  But tests 12,13, and
especially 14 (LU decomposition) are faster.  Even I/O seems similar, which
probably means I was comparing apples and oranges, since my original result
(around .45 seconds) was obtained using a different version of Linux and the C
libraries (I was forced to upgrade in order to use 5.3).  When compared side by
side in this way, results really are quite similar.  The geometric means were
0.35871513 (v5.3) and 0.38800894 (v5.2.1), indicating the *slight* speedup of
5.3 on specific tests, most notably LU decomp.  

IDLSpecMark's were 0.873758 (v5.3) versus 0.891246 (v5.2.1), but had a good deal
of variability on subsequent invocations, and were certainly consistent with
similar performance.  My suggestion is to look to the operating system, or try
installing 5.2.1 on your new machine and comparing that way.

JD


-- 
 J.D. Smith                             |*|      WORK: (607) 255-5842    
 Cornell University Dept. of Astronomy  |*|            (607) 255-6263
 304 Space Sciences Bldg.               |*|       FAX: (607) 255-5875 
 Ithaca, NY 14853                       |*|